| 1 | //===- PPCBoolRetToInt.cpp ------------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements converting i1 values to i32/i64 if they could be more |
| 10 | // profitably allocated as GPRs rather than CRs. This pass will become totally |
| 11 | // unnecessary if Register Bank Allocation and Global Instruction Selection ever |
| 12 | // go upstream. |
| 13 | // |
| 14 | // Presently, the pass converts i1 Constants, and Arguments to i32/i64 if the |
| 15 | // transitive closure of their uses includes only PHINodes, CallInsts, and |
| 16 | // ReturnInsts. The rational is that arguments are generally passed and returned |
| 17 | // in GPRs rather than CRs, so casting them to i32/i64 at the LLVM IR level will |
| 18 | // actually save casts at the Machine Instruction level. |
| 19 | // |
| 20 | // It might be useful to expand this pass to add bit-wise operations to the list |
| 21 | // of safe transitive closure types. Also, we miss some opportunities when LLVM |
| 22 | // represents logical AND and OR operations with control flow rather than data |
| 23 | // flow. For example by lowering the expression: return (A && B && C) |
| 24 | // |
| 25 | // as: return A ? true : B && C. |
| 26 | // |
| 27 | // There's code in SimplifyCFG that code be used to turn control flow in data |
| 28 | // flow using SelectInsts. Selects are slow on some architectures (P7/P8), so |
| 29 | // this probably isn't good in general, but for the special case of i1, the |
| 30 | // Selects could be further lowered to bit operations that are fast everywhere. |
| 31 | // |
| 32 | //===----------------------------------------------------------------------===// |
| 33 | |
| 34 | #include "PPC.h" |
| 35 | #include "PPCTargetMachine.h" |
| 36 | #include "llvm/ADT/DenseMap.h" |
| 37 | #include "llvm/ADT/STLExtras.h" |
| 38 | #include "llvm/ADT/SmallPtrSet.h" |
| 39 | #include "llvm/ADT/SmallVector.h" |
| 40 | #include "llvm/ADT/Statistic.h" |
| 41 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 42 | #include "llvm/IR/Argument.h" |
| 43 | #include "llvm/IR/Dominators.h" |
| 44 | #include "llvm/IR/Function.h" |
| 45 | #include "llvm/IR/IRBuilder.h" |
| 46 | #include "llvm/IR/Instruction.h" |
| 47 | #include "llvm/IR/Instructions.h" |
| 48 | #include "llvm/IR/Type.h" |
| 49 | #include "llvm/IR/Use.h" |
| 50 | #include "llvm/IR/User.h" |
| 51 | #include "llvm/IR/Value.h" |
| 52 | #include "llvm/Pass.h" |
| 53 | #include "llvm/Support/Casting.h" |
| 54 | #include <cassert> |
| 55 | |
| 56 | using namespace llvm; |
| 57 | |
| 58 | namespace { |
| 59 | |
| 60 | #define DEBUG_TYPE "ppc-bool-ret-to-int" |
| 61 | |
| 62 | STATISTIC(NumBoolRetPromotion, |
| 63 | "Number of times a bool feeding a RetInst was promoted to an int" ); |
| 64 | STATISTIC(NumBoolCallPromotion, |
| 65 | "Number of times a bool feeding a CallInst was promoted to an int" ); |
| 66 | STATISTIC(NumBoolToIntPromotion, |
| 67 | "Total number of times a bool was promoted to an int" ); |
| 68 | |
| 69 | class PPCBoolRetToInt : public FunctionPass { |
| 70 | static SmallPtrSet<Value *, 8> findAllDefs(Value *V) { |
| 71 | SmallPtrSet<Value *, 8> Defs; |
| 72 | SmallVector<Value *, 8> WorkList; |
| 73 | WorkList.push_back(Elt: V); |
| 74 | Defs.insert(Ptr: V); |
| 75 | while (!WorkList.empty()) { |
| 76 | Value *Curr = WorkList.pop_back_val(); |
| 77 | auto *CurrUser = dyn_cast<User>(Val: Curr); |
| 78 | // Operands of CallInst/Constant are skipped because they may not be Bool |
| 79 | // type. For CallInst, their positions are defined by ABI. |
| 80 | if (CurrUser && !isa<CallInst>(Val: Curr) && !isa<Constant>(Val: Curr)) |
| 81 | for (auto &Op : CurrUser->operands()) |
| 82 | if (Defs.insert(Ptr: Op).second) |
| 83 | WorkList.push_back(Elt: Op); |
| 84 | } |
| 85 | return Defs; |
| 86 | } |
| 87 | |
| 88 | // Translate a i1 value to an equivalent i32/i64 value: |
| 89 | Value *translate(Value *V) { |
| 90 | assert(V->getType() == Type::getInt1Ty(V->getContext()) && |
| 91 | "Expect an i1 value" ); |
| 92 | |
| 93 | Type *IntTy = ST->isPPC64() ? Type::getInt64Ty(C&: V->getContext()) |
| 94 | : Type::getInt32Ty(C&: V->getContext()); |
| 95 | |
| 96 | if (auto *P = dyn_cast<PHINode>(Val: V)) { |
| 97 | // Temporarily set the operands to 0. We'll fix this later in |
| 98 | // runOnUse. |
| 99 | Value *Zero = Constant::getNullValue(Ty: IntTy); |
| 100 | PHINode *Q = |
| 101 | PHINode::Create(Ty: IntTy, NumReservedValues: P->getNumIncomingValues(), NameStr: P->getName(), InsertBefore: P->getIterator()); |
| 102 | for (unsigned I = 0; I < P->getNumOperands(); ++I) |
| 103 | Q->addIncoming(V: Zero, BB: P->getIncomingBlock(i: I)); |
| 104 | return Q; |
| 105 | } |
| 106 | |
| 107 | IRBuilder IRB(V->getContext()); |
| 108 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
| 109 | IRB.SetInsertPoint(I->getNextNode()); |
| 110 | else |
| 111 | IRB.SetInsertPoint(TheBB: &Func->getEntryBlock(), IP: Func->getEntryBlock().begin()); |
| 112 | return IRB.CreateZExt(V, DestTy: IntTy); |
| 113 | } |
| 114 | |
| 115 | typedef SmallPtrSet<const PHINode *, 8> PHINodeSet; |
| 116 | |
| 117 | // A PHINode is Promotable if: |
| 118 | // 1. Its type is i1 AND |
| 119 | // 2. All of its uses are ReturnInt, CallInst, or PHINode |
| 120 | // AND |
| 121 | // 3. All of its operands are Constant or Argument or |
| 122 | // CallInst or PHINode AND |
| 123 | // 4. All of its PHINode uses are Promotable AND |
| 124 | // 5. All of its PHINode operands are Promotable |
| 125 | static PHINodeSet getPromotablePHINodes(const Function &F) { |
| 126 | PHINodeSet Promotable; |
| 127 | // Condition 1 |
| 128 | for (auto &BB : F) |
| 129 | for (auto &I : BB) |
| 130 | if (const auto *P = dyn_cast<PHINode>(Val: &I)) |
| 131 | if (P->getType()->isIntegerTy(Bitwidth: 1)) |
| 132 | Promotable.insert(Ptr: P); |
| 133 | |
| 134 | SmallVector<const PHINode *, 8> ToRemove; |
| 135 | for (const PHINode *P : Promotable) { |
| 136 | // Condition 2 and 3 |
| 137 | auto IsValidUser = [] (const Value *V) -> bool { |
| 138 | return isa<ReturnInst>(Val: V) || isa<CallInst>(Val: V) || isa<PHINode>(Val: V); |
| 139 | }; |
| 140 | auto IsValidOperand = [] (const Value *V) -> bool { |
| 141 | return isa<Constant>(Val: V) || isa<Argument>(Val: V) || isa<CallInst>(Val: V) || |
| 142 | isa<PHINode>(Val: V); |
| 143 | }; |
| 144 | const auto &Users = P->users(); |
| 145 | const auto &Operands = P->operands(); |
| 146 | if (!llvm::all_of(Range: Users, P: IsValidUser) || |
| 147 | !llvm::all_of(Range: Operands, P: IsValidOperand)) |
| 148 | ToRemove.push_back(Elt: P); |
| 149 | } |
| 150 | |
| 151 | // Iterate to convergence |
| 152 | auto IsPromotable = [&Promotable] (const Value *V) -> bool { |
| 153 | const auto *Phi = dyn_cast<PHINode>(Val: V); |
| 154 | return !Phi || Promotable.count(Ptr: Phi); |
| 155 | }; |
| 156 | while (!ToRemove.empty()) { |
| 157 | for (auto &User : ToRemove) |
| 158 | Promotable.erase(Ptr: User); |
| 159 | ToRemove.clear(); |
| 160 | |
| 161 | for (const PHINode *P : Promotable) { |
| 162 | // Condition 4 and 5 |
| 163 | const auto &Users = P->users(); |
| 164 | const auto &Operands = P->operands(); |
| 165 | if (!llvm::all_of(Range: Users, P: IsPromotable) || |
| 166 | !llvm::all_of(Range: Operands, P: IsPromotable)) |
| 167 | ToRemove.push_back(Elt: P); |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | return Promotable; |
| 172 | } |
| 173 | |
| 174 | typedef DenseMap<Value *, Value *> B2IMap; |
| 175 | |
| 176 | public: |
| 177 | static char ID; |
| 178 | |
| 179 | PPCBoolRetToInt() : FunctionPass(ID) {} |
| 180 | |
| 181 | bool runOnFunction(Function &F) override { |
| 182 | if (skipFunction(F)) |
| 183 | return false; |
| 184 | |
| 185 | auto *TPC = getAnalysisIfAvailable<TargetPassConfig>(); |
| 186 | if (!TPC) |
| 187 | return false; |
| 188 | |
| 189 | auto &TM = TPC->getTM<PPCTargetMachine>(); |
| 190 | ST = TM.getSubtargetImpl(F); |
| 191 | Func = &F; |
| 192 | |
| 193 | PHINodeSet PromotablePHINodes = getPromotablePHINodes(F); |
| 194 | B2IMap Bool2IntMap; |
| 195 | bool Changed = false; |
| 196 | for (auto &BB : F) { |
| 197 | for (auto &I : BB) { |
| 198 | if (auto *R = dyn_cast<ReturnInst>(Val: &I)) |
| 199 | if (F.getReturnType()->isIntegerTy(Bitwidth: 1)) |
| 200 | Changed |= |
| 201 | runOnUse(U&: R->getOperandUse(i: 0), PromotablePHINodes, BoolToIntMap&: Bool2IntMap); |
| 202 | |
| 203 | if (auto *CI = dyn_cast<CallInst>(Val: &I)) |
| 204 | for (auto &U : CI->operands()) |
| 205 | if (U->getType()->isIntegerTy(Bitwidth: 1)) |
| 206 | Changed |= runOnUse(U, PromotablePHINodes, BoolToIntMap&: Bool2IntMap); |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | return Changed; |
| 211 | } |
| 212 | |
| 213 | bool runOnUse(Use &U, const PHINodeSet &PromotablePHINodes, |
| 214 | B2IMap &BoolToIntMap) { |
| 215 | auto Defs = findAllDefs(V: U); |
| 216 | |
| 217 | // If the values are all Constants or Arguments, don't bother |
| 218 | if (llvm::none_of(Range&: Defs, P: [](Value *V) { return isa<Instruction>(Val: V); })) |
| 219 | return false; |
| 220 | |
| 221 | // Presently, we only know how to handle PHINode, Constant, Arguments and |
| 222 | // CallInst. Potentially, bitwise operations (AND, OR, XOR, NOT) and sign |
| 223 | // extension could also be handled in the future. |
| 224 | for (Value *V : Defs) |
| 225 | if (!isa<PHINode>(Val: V) && !isa<Constant>(Val: V) && |
| 226 | !isa<Argument>(Val: V) && !isa<CallInst>(Val: V)) |
| 227 | return false; |
| 228 | |
| 229 | for (Value *V : Defs) |
| 230 | if (const auto *P = dyn_cast<PHINode>(Val: V)) |
| 231 | if (!PromotablePHINodes.count(Ptr: P)) |
| 232 | return false; |
| 233 | |
| 234 | if (isa<ReturnInst>(Val: U.getUser())) |
| 235 | ++NumBoolRetPromotion; |
| 236 | if (isa<CallInst>(Val: U.getUser())) |
| 237 | ++NumBoolCallPromotion; |
| 238 | ++NumBoolToIntPromotion; |
| 239 | |
| 240 | for (Value *V : Defs) { |
| 241 | auto [It, Inserted] = BoolToIntMap.try_emplace(Key: V); |
| 242 | if (Inserted) |
| 243 | It->second = translate(V); |
| 244 | } |
| 245 | |
| 246 | // Replace the operands of the translated instructions. They were set to |
| 247 | // zero in the translate function. |
| 248 | for (auto &Pair : BoolToIntMap) { |
| 249 | auto *First = dyn_cast<User>(Val: Pair.first); |
| 250 | auto *Second = dyn_cast<User>(Val: Pair.second); |
| 251 | assert((!First || Second) && "translated from user to non-user!?" ); |
| 252 | // Operands of CallInst/Constant are skipped because they may not be Bool |
| 253 | // type. For CallInst, their positions are defined by ABI. |
| 254 | if (First && !isa<CallInst>(Val: First) && !isa<Constant>(Val: First)) |
| 255 | for (unsigned I = 0; I < First->getNumOperands(); ++I) |
| 256 | Second->setOperand(i: I, Val: BoolToIntMap[First->getOperand(i: I)]); |
| 257 | } |
| 258 | |
| 259 | Value *IntRetVal = BoolToIntMap[U]; |
| 260 | Type *Int1Ty = Type::getInt1Ty(C&: U->getContext()); |
| 261 | auto *I = cast<Instruction>(Val: U.getUser()); |
| 262 | Value *BackToBool = |
| 263 | new TruncInst(IntRetVal, Int1Ty, "backToBool" , I->getIterator()); |
| 264 | U.set(BackToBool); |
| 265 | |
| 266 | return true; |
| 267 | } |
| 268 | |
| 269 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 270 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 271 | FunctionPass::getAnalysisUsage(AU); |
| 272 | } |
| 273 | |
| 274 | private: |
| 275 | const PPCSubtarget *ST; |
| 276 | Function *Func; |
| 277 | }; |
| 278 | |
| 279 | } // end anonymous namespace |
| 280 | |
| 281 | char PPCBoolRetToInt::ID = 0; |
| 282 | INITIALIZE_PASS(PPCBoolRetToInt, "ppc-bool-ret-to-int" , |
| 283 | "Convert i1 constants to i32/i64 if they are returned" , false, |
| 284 | false) |
| 285 | |
| 286 | FunctionPass *llvm::createPPCBoolRetToIntPass() { return new PPCBoolRetToInt(); } |
| 287 | |